Despite the growing use of population viability analysis (PTA), the predict
ions of these models rarely have been tested with field data that were not
used In initially developing the model, We review and discuss a suite of me
thods that may be used to test the predictive ability of models used in PVA
. In addition to testing mean predictions, appropriate methods must analyze
the probability distribution of the model predictions. The methods we disc
uss provide tests of the mean predictions, the predicted frequency, of even
ts such cis extinction and colonization, and the predicted probability dist
ribution of state variables. We discuss visual approaches based on plots of
observations versus the predictions and statistical approaches based on de
termining significant differences between observations and predictions. The
advantages and disadvantages of each method are identified. The best metho
ds test the statistical distribution of the predictions; those that ignore
variability are meaningless. Although ive recognize that the quality of a m
odel is not solely a function of its predictive abilities, tests help reduc
e inherent model uncertainty. The role of model testing is not to prove the
truth of a model, which is impossible because models are never a perfect d
escription of reality. Rather. testing should help identify the weakest asp
ects of models so they can be improved. We provide a framework for using mo
del testing to improve the predictive performance of PVA models, through an
iterative process of model development, testing, subsequent modification a
nd re-testing.